Pip: Welcome to The International Telegraph — where this week, the future of education, software pricing, and emotional AI all apparently arrived at once and immediately started arguing.
Mara: Jay Davis has been tracking all of it: a private school network betting big on AI-powered learning, the murky billing math behind vibe coding platforms, and China's new rules forcing major tech companies to pull their companion AI features. Let's start with the school that thinks two hours of AI a day is enough.
Alpha School's Bet on AI Learning
Mara: Alpha School is planning to expand from around a dozen campuses to roughly 50 this school year, adding markets from Atlanta to Boston to Denver, with its New York City campus alone expected to grow from 22 to 160 students.
Pip: The model is unusual. Students cap AI time at two hours a day for core academics, then spend the rest of the day on life-skills workshops — and the adults in the room are not licensed teachers but paid coaches.
Mara: Price told Axios directly: "AI allows us to finally, truly transform the educational system and not just have it be a little tool that keeps a traditional, one-teacher-in-front-of-a-classroom model going."
Pip: The school claims students grow 2.6 times faster than peers on standardized MAP tests — but Stanford professor Victor Lee told CNN that Alpha has refused independent research access to verify those claims, which is a choice that tends to raise more questions than it answers.
Mara: Penn State professor Gerald LeTendre put the broader concern plainly to Axios: "Right now we're conducting a gigantic experiment on the human population, on our children, without any good scientific knowledge." Pennsylvania also denied a cyber charter application from Unbound Academy, the affiliated school using the same two-hour model, calling it untested.
Pip: The regulatory picture gets more complicated from there — and so does the bill. Which brings us to what vibe coding actually costs.
The Hidden Bill Inside Vibe Coding
Mara: The vibe coding segment is really about a structural question: when an AI coding platform sells you credits, what exactly are you buying, and who benefits when the AI gets it wrong?
Pip: Cursor's public apology after its June 2025 pricing change is the clearest statement of how badly this can go. The company wrote: "Our recent pricing changes for individual plans were not communicated clearly, and we take full responsibility" — and offered refunds for unexpected charges over a three-week window.
Mara: So the upshot is that users were being charged in ways the platform itself admitted were not intuitive. The terminology problem was real: Cursor acknowledged it had described usage limits as "rate limits" rather than a credit pool, and that "unlimited usage" applied only to one model tier, not all of them.
Pip: Replit had its own version of this. After the Agent 3 launch, The Register reported one user spent a thousand dollars in a single week — against typical monthly costs of a hundred and eighty to two hundred dollars.
Mara: The underlying dynamic, which the piece on vibe coding billing explains carefully, is that editing existing code triggers multiple sub-agents to review, plan, check, and fix — each pass re-reading the whole codebase. Bolt's own documentation confirms it: most token usage comes from syncing the project's file system to the AI, and the larger the project, the more tokens every message consumes.
Pip: The bring-your-own-key alternative, used by tools like Cline, removes the platform from the token transaction entirely — you pay Anthropic's published rates directly. But removing the markup does not remove the cost, because the context problem is the same regardless.
Mara: The piece also notes a structural conflict worth naming: on a credit model, every failed build is revenue for the platform. The incentive to write good code the first time does not sit with the platform — it sits against it.
Pip: The pricing model, it turns out, tells you whose side the platform is on. That same question — who controls the system and who bears the risk — runs through what China just decided about companion AI.
China Shuts Down Companion AI
Mara: China's new rules on anthropomorphic AI services are not a vague policy signal — they are a hard deadline with named companies already pulling features to comply.
Pip: ByteDance and Alibaba are the headline cases. Doubao notified users that its agent feature would go offline on July 15 because of "product function adjustments," with agent data becoming inaccessible after October 15. Qwen moved faster, disabling humanlike interactive agents on July 10, ahead of the formal deadline.
Mara: The regulation itself, the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, targets services that "simulate human personality traits, thinking patterns and communication styles to provide sustained emotional interaction." Five agencies issued it jointly, including the Cyberspace Administration of China, and it took effect July 15, 2026.
Pip: Tencent's Yuanbao had already pulled a comparable feature in June, before the deadline arrived — which suggests the industry read the direction of travel clearly enough not to wait.
Mara: What the rules require is specific. Artificial Intelligence News reported that companion or family-member AI services for minors are prohibited outright, and guardian consent is required for users under 14. Platforms must build in anti-addiction systems, usage notifications, and exit mechanisms, and they are prohibited from engineering emotional dependence or using emotional manipulation to drive decisions.
Pip: Quartz adds that companies cannot use private conversations to train their models, and that any AI behavior creating attachments strong enough to replace real relationships is banned. That is a genuinely difficult line to draw in practice, but the regulation draws it anyway.
Mara: The rules do carve out space for customer service bots, workplace assistants, and educational tools — provided those tools are not designed for sustained emotional engagement. The distinction the regulation is trying to enforce is between a tool and a relationship.
Pip: Which is, incidentally, the same distinction Alpha School is navigating on the other side of the world — between AI as a tutor and AI as something more.
Mara: Three different domains this week, and the same question underneath all of them: when a system is designed to engage you, who is responsible for what happens next?
Pip: Regulators in Beijing answered that. Educators in Austin are still working it out. And somewhere, a vibe coder is watching their credit balance fall and wondering the same thing.
